earhian / humpback-whale-identification-1st- Goto Github PK

Hello
Are the pretrained models available to download?
Thanks

as below

老哥你如何做的mask

FileNotFoundError: File b'./input/model_50A_slim_ensemble.csv' does not exist

您好 我们想在您的基础上训练一个分类网络，但细节的地方有两个位置不太理解，能不能麻烦解释一下，万分感谢。
第一个是results为什么后面还要torch.cat一个torch.ones_like
results = torch.cat([torch.sigmoid(results), torch.ones_like(results[:, :1]).float().cuda() * 0.5], 1)
第二个是这里为什么要乘以16
out = self.fc(global_feat) * 16

My name is Taiyo Suzuki. I live in Yokohama, Japan, and I speak English.
I'm studying data science to seriously accelerate the Singularity.
I want you to teach me kaggle.
I don't know how to build a model.
I'm an intern with no major in machine learning.
Thank you for your time.

flake8 testing of https://github.com/earhian/Humpback-Whale-Identification-1st- on Python 3.7.1

$ flake8 . --count --select=E901,E999,F821,F822,F823 --show-source --statistics

./models/modelZoo/convert_from_mxnet.py:111:5: F821 undefined name 'main'
    main()    ^
1     F821 undefined name 'main'
1

E901,E999,F821,F822,F823 are the "showstopper" flake8 issues that can halt the runtime with a SyntaxError, NameError, etc. These 5 are different from most other flake8 issues which are merely "style violations" -- useful for readability but they do not effect runtime safety.

• F821: undefined name name
• F822: undefined name name in __all__
• F823: local variable name referenced before assignment
• E901: SyntaxError or IndentationError
• E999: SyntaxError -- failed to compile a file into an Abstract Syntax Tree

Ran into two problems.

bboxs.csv file not found. I am assuming that the bbox_by_masks.csv is the correct bboxs.csv file

Cuda Out of Memory Error.
Having 2 1080Ti GPU with 11G memory each. Got the OOM error as following.

Traceback (most recent call last):
  File "train.py", line 310, in <module>
    train(freeze, fold_index, model_name, min_num_class, checkPoint_start, lr, batch_size)
  File "train.py", line 261, in train
    global_feat, local_feat, results = data_parallel(model,images)

changing batch_size from 12 to 6 doesn't help to solve this. Would you like to share your training environment. Thanks.

Hello, perfect job! But I wonder how did you get the bounding boxes? Is there a source code by which I can regenerate this .csv file?

Can you explain more exactly what the mask is? THANKS.

I want use your triplet_loss function to train a classification model with imagenet, and i got a mistake at triplet_loss.py line 70
dist_ap, relative_p_inds = torch.max((dist_mat * is_pos.float()).contiguous().view(N, -1), 1,keepdim=True)
if my label is [1, 1, 2, 3] then N=4 i got
RuntimeError: shape '[4, -1]' is invalid for input of size 6
if my label is [1, 1, 2, 2] and it worked
Is my understanding correct?Do I need to generate labels in this form[1,1,2,2]?

earhian, Thanks for sharing !!!

i'd like to train your model..

could you use pretrained SENET model?

if I use pretrained model, there is input channel size problem(your model: 4, imagenet: 3) in load state_dict().

How could I fix It?

我很困惑arcface的实现过程，为为什么要在越界使用cosface

'senet154' model pretrained on imagenet needs input_channel=3, but your model's input is 4. And I can't find the transformation between the input channel.How you do that ?Thanks!

你好，我在学习你写的代码的时候，对于你的一段在local_dist()函数中的代码产生了疑惑。你使用了这个函数，直接计算了两个三维矩阵的距离。我理解中的做法，是在实际使用过程中，x和y是同一个矩阵，其维数相同，把这两个矩阵给做一个按实例样本数做第一维，剩余的做flatten之后求其欧拉距离或者余弦距离。但是你的做法是把通道数最后一维， 把其的flatten，求欧拉距离之后，再使用了一个 (exp(x) - 1) / (exp(x) + 1)， 这个函数的作用是用作把值归到[0, 1]的么，这种的归一化我还没有见过，这是某种特殊的作法么？其次你把距离矩阵的维做了交换之后使用了一次动态规划来求一个距离矩阵，那这意思可能是之前在做的都是在求一个关于距离长度度量的一个矩阵，我对此做法感到很不解。我也在网上找了一些资料，但是没有找到类似做法的解释。如果可以，你可以给我一些帮助么。感谢老铁， 你的这个代码写的很棒。